Gradient-boosted signals on real market data. Triple-barrier labels, purged K-Fold CV, walk-forward backtests. Verified by Deflated Sharpe and Probability of Backtest Overfitting.
Triple-barrier labels (López de Prado). Gradient-boosted trees trained with purged K-Fold CV + embargo so future bars never leak into the past. Folds average into a single conviction score per name. No deep learning, no transformers — just disciplined gradient boosting.
Every submission runs through a layered gate — kill-switch, position caps, drawdown kill, max-positions — before the broker ever hears about it. The terminal simply cannot bypass it.
Kill-switch is a single Redis key, persisted with the reason and the timestamp. Gauged in Prometheus. Read on every submit, never cached. It's the big red button that actually works.
Every published result ships with a reproducibility manifest: git sha, config hash, data fingerprint, environment snapshot. And its performance is adjusted for how many trials were burned finding it.
Kill-switch · position caps · drawdown · max-positions.
Pending → submitted → filled / cancelled / rejected. Idempotent.
Top-K per universe, ordered by rank then confidence.
Reconnecting WS consumer, Redis pub/sub fan-out, 15 s heartbeats.
Equal-weight top-K, bps cost model, Sharpe + drawdown + DSR.
11 APIs keyed. Canonical Postgres. Zero synthetic paths.
No fake data.
No synthetic paths.
Real providers. Real markets.
Verified by deflated Sharpe.
Or it does not ship.
Measured from the commit log, not the pitch deck.
Real BUY/HOLD/SELL from LightGBM trained on 718k rowsof real Alpaca S&P 500 data, 2018–2026. Sorted by model conviction. Zero synthetic paths.
